Olmo 3 represents a significant step in making large language model (LLM) research more accessible. Unlike many "open" models that only share weights, Olmo 3 provides a complete set of training artifacts, including data, code, and checkpoints. This transparency allows anyone to retrain the model from scratch. Olmo 3 comes in two sizes: 7 billion and 32 billion parameters, with the 32B version being labeled as the strongest fully open reasoning model to date. While Olmo 3 doesn't quite match the performance of leading closed models, it excels in transparency and usability for researchers. The training process is detailed, with a three-stage approach: general pretraining on a large corpus, midtraining using targeted high-quality data, and a context extension phase. It also utilizes advanced training infrastructure, leveraging 1,024 NVIDIA H100 GPUs and a Fully-Sharded Data Parallel (FSDP) system to optimize memory and processing efficiency during training. The article outlines different model variants derived from Olmo 3. The Instruct models focus on quick responses and instruction following, while the Think models are tailored for complex reasoning by generating detailed thought processes. The RL-Zero models apply reinforcement learning directly on the pretrained base model. This structured approach to training and the release of comprehensive resources positions Olmo 3 as a valuable tool for those looking to contribute to open LLM research.